In that Mr. Elon said

AI has not made an impact yet on productivity to increase good and services faster than money supply .

Elon Musk: A Different Conversation w/ Nikhil Kamath (49:57- 50:22)

So i got question what impact AI has does on productivity ?

As per current state , there are developers who are using chatgpt for the code generation to save time and also getting headache to understand what is been generated ….

there are content writers who use it for paraphrasing and getting ideas , copywriting and lot more

there are every day people who take feedback from chatgpt

Tip : if you want to get actual feedback from AI Models like chatgpt, claude,grok . just write “Be Brutally Honest” while prompting .it better for reinforcement learning.

coming to our curious question . we all are dealing with one thing that takes our lot of time .

EMAIL

everyday i see my inbox messy and lot of unrelated email that i focus on .majorly no of irrelevant cold email that we receive everyday . currently spam filter can’t catch them .so i have to individual read the title or check the email body and recognize as spam and then delete them or block them .

i have seen good implementation of AI that saves time for your mailbox .

A tool that does 3 things

• a personal assistant that reads your mail on your behalf and categorize the mail and auto archive – so that you only focus what is essential (organizing your mailbox) Don’t worry it will not archive the email that you have communicate previously or anyone within your organization

• It reminds you to follow up if there is no reply from sender and draft the email based on your writing tone by reading previous mail . (useful for leads ,sales and marketing team)

• you can write your own category and tell about category via prompting . it categorize your mail by reading all mail and also do it for upcoming email. (priorities)

this are very small thing but makes big different when we see the compounding impact everyday .the tool name is “Supersmart.ai mail”. (This is not a sponsored post)

But a simple way to tell you guy what ai can solve how much it can solve …just to give you reality check … i will bring different usecase across different domain and add new usecase each week . you can suggest your usage with ai which you use in daily life i will publish those … to help everyone learn from it .

if you want to know how it happens at a backend .

Currently in AI domain, design is the most crucial thing . if you look chatgpt or perplexity it is been design such a way that it can help you find information . the way google do it one time search and again you have to change the search query if you are unsatisfied .as a search it is a iterative process until user is not satisfied which perplexity has overcome it through its design.

Mr. – Rahul Vohra

coming to the current usecase , understanding superhuman.ai tool which has founded by Rahul Vohra. Superhuman’s architecture reveals sophisticated systems design decisions around embedding models, inference optimization, and offline-first web application engineering.

it a email client software . It majorly creating a system that:

• Filters out all of the noise so you can stay focused
• Sorts messages by genuine priority
• Presents information in digestible summaries
• Drafts contextually appropriate responses

for this Major functionality

• Email Classification – Categorize incoming emails
• Smart Reply Generation – Suggest contextually appropriate replies
• offline capabilities – work without internet ; sync when reconnected
• Real time performance – supports millions of users across multiple time zones
• Global scale – Supports millions of users across multiple time zones

Non – Functional :

• Throughput: 500M+ emails/day globally . It can expect this many email in a day overall
• Latency: P95 < 500ms for email classification, P99 < 1s . email classification in the has to be less than 500 milliseconds . if it exceed more than 2.5 second the user experience gets worst and user have to wait for it .
• Availability: 99.99% uptime
• Cost Efficiency: <$0.001 per email processed
• Model Accuracy: >95% precision on high-confidence predictions .

how does AI System work ?

Photo by Joshua J. Cotten on Unsplash

Custom Models over Generic Models

Rather than relying on generic model . Superhuman have deployed a suite a dozens of custom model (fine-tuned embedding models optimized for specific email classification task. as generic model provides lower accuracy for specialized email task . Superhuman uses a approach “Superhuman calibration” – a technique where they identify a subset of data on which model achieve great performance then automate those cases.

Real Example:

For email classification (spam, pitch, marketing, news, calendar, important), a standard classifier might have:

• 1 in 10 error rate (90% accuracy) — too low for production
• Human error rate: 1 in 50 (98% accuracy)

Instead of deploying the 90% model to all emails, Superhuman calibrates it to identify when confidence is high (e.g., clearly promotional newsletters). Only those high-confidence predictions are automated. Lower-confidence cases are flagged for human review or manual categorization.

Feature Engineering & Processing :

It uses NLP preprocessing but with domain specific optimization :

1. Tokenization: Break emails into tokens
2. Text Normalization: Lowercasing, removing special characters (but preserve email-specific signals like subject/body distinction)
3. Stemming/Lemmatization: Reduce words to root form
4. Word Embeddings: Convert text to vectors

Key Features Extracted:

• Subject line content (high weight for importance signals)
• Sender reputation (previous interaction history with sender)
• Email body length (very short = likely promotional)
• Presence of links/attachments (signals intent)
• Email metadata (From, To, CC, Date patterns)
• Temporal patterns (send time, reply speed)

Multi-Task Learning Architecture

Photo by Ismail Salad Osman Hajji dirir on Unsplash

Based on technical patterns used in email systems, Superhuman likely implements multi-task learning:

Task 1: Importance Classification

• Binary: {Important, Not Important}
• Trained on user labels (starred emails, replied-to emails)

Task 2: Category Classification

• Multi-class: {Pitch, Marketing, News, Calendar, Other, Personal}
• Split Inbox uses these predictions to auto-organize

Task 3: Spam Detection

• Binary classifier with extremely high precision (false positives unacceptable)

Shared Representation:

All tasks share a common encoder (embedding layer), allowing transfer learning. This means:

• Spam detection knowledge helps with “Pitch” classification
• Importance signals inform category labeling
• One model improves multiple outputs simultaneously

Training Data & Continuous Learning

Data Collection:

• User actions (starred/archived/replied emails) generate weak labels
• User feedback on AI suggestions refines models
• Team collaboration signals (replied vs. ignored) provide importance labels

Continuous Learning:

System doesn’t retrain daily. Instead, they likely use:

• Monthly full retraining cycles
• A/B testing of model variations on user subsets
• Feedback loops: user actions → model improvement

Inference Infrastructure :

they need to process 500+ email message globally . Every email need to classify and need to provide instant feedback to users . They used Baseten Embeddings Inference which uses

• TensortRT LLM Runtime – NVIDIA’s optimized inference engine specifically for embeddings and classifiers . it implements kernel fusion for embedding operations .Custom CUDA kernel for attention free architecture
• Multi Cloud Management – Latency aware route across near GPU cluster , no single point of failure (active-active reliability ) , auto scale based on request volume .
• Performance Client – Batching optimization (Group multiple embedding requests) ,connection pooling to reduce overhead ,eliminate network bottleneck in the inference path Results – 80% latency reduction (P95 latency dropped from 2.5s to 500ms) Deployed dozens of model via unified deployed interface .

Performance Metrics & Monitoring

1. Inference Latency:
   • P50, P95, P99 latencies for each model
   • Target: <500ms for email classification
   • Alert if P95 exceeds 750ms
2. Model Accuracy:
   • Precision/Recall for each classification task
   • False positive rate (users prefer missing important email alerts vs. annoying false positives)
   • A/B testing new models on 5% of users before full rollout
3. Business Metrics:
   • Hours saved per user (target: 4+ hours/week)
   • User retention (how many users keep using Superhuman after 30 days trial)
   • Engagement (emails replied to via AI suggestions)

Technical Decisions Worth Learning From

• Custom Models over Closed-Source
• Superhuman Calibration Pattern
• Baseten for Inference
• Modifier Architecture (for offline data storage mechanism)

Superhuman’s success isn’t about using the most advanced models—it’s about architectural decisions optimized for latency, accuracy, and offline resilience. The “Superhuman Calibration” pattern (deploy only high-confidence predictions) is applicable to any production ML system. Their infrastructure choices (Baseten over in-house) reflect a pragmatic focus.

Soon this feature will be establish as inbuild functionality in most of the popular email management tool .

Thank you for reading !! Share with someone who want to know real use of AI.

Thanks you for reading UsecaseinAI Substack!